OpenAI fires three safety researchers for "mishandling research information"

If you are an AI safety researcher, your ability to collaborate with external experts on critical safety issues without fear of reprisal is now compromised.

OpenAI has fired three safety researchers: Jasmine Wang, Tomek Korbak, and Mikita Balesni. The company stated they were dismissed for mishandling sensitive company information and violating policies by sharing it with a third-party AI safety organization. The researchers, in an open letter, dispute these claims, asserting that their actions were within OpenAI's norms at the time and necessary for effective AI safety collaboration with external experts. They argue their dismissals signal a chilling effect on OpenAI's culture, discouraging open dialogue and safety concerns. The researchers deny leaking information about less monitorable architectures to The Information, and Wang specifically clarified her situation involved an executive's email she had access to for recruiting, which she reported when a sensitive email was mistakenly opened. OpenAI maintains the firings were not retaliatory and were due to a pattern of misconduct, not for raising safety concerns.

The dismissal of these prominent safety researchers suggests a significant tension between OpenAI's stated commitment to open culture and safety, and its internal management of sensitive information and external collaborations. This event raises questions about the company's risk tolerance and its approach to managing dissenting voices or researchers who push the boundaries of information sharing for safety's sake. It could lead to increased scrutiny from both internal employees and external regulatory bodies regarding OpenAI's safety protocols and transparency.

This incident may prompt other AI labs to re-evaluate their own policies on information sharing and external collaboration among safety teams, potentially leading to more restrictive guidelines. The researchers' warning of a "chilling effect" could impact the willingness of AI safety experts to voice concerns or engage in the open dialogue crucial for developing AGI responsibly. This could ultimately slow down the pace of crucial safety research and external accountability in the rapidly advancing field of artificial intelligence.